onnxruntime update to 1.1.0, before convert, onnxruntime will run model to...

onnxruntime update to 1.1.0, before convert, onnxruntime will run model to check the correctness of model

README.md

@@ -15,7 +15,7 @@ paddlepaddle >= 1.6.0
 **按需安装以下依赖**  
 tensorflow ： tensorflow == 1.14.0  
 caffe ： 无  
-onnx ： onnx == 1.5.0  onnxruntime == 0.4.0
+onnx ： onnx == 1.6.0  onnxruntime == 1.1.0
 
 ## 安装
 ### 安装方式一（推荐）

setup.py

@@ -26,6 +26,6 @@ setuptools.setup(
     entry_points={
         'console_scripts': [
             'x2paddle=x2paddle.convert:main',
-            'onnx_infer=x2paddle.onnx_infer:main'
+            #            'onnx_infer=x2paddle.onnx_infer:main'
         ]
     })

x2paddle/convert.py

@@ -155,18 +155,18 @@ def onnx2paddle(model_path, save_dir, params_merge=False):
     try:
         import onnx
         version = onnx.version.version
-        if version != '1.5.0':
-            print("onnx==1.5.0 is required")
+        if version != '1.6.0':
+            print("onnx==1.6.0 is required")
             return
     except:
-        print("onnx is not installed, use \"pip install onnx==1.5.0\".")
+        print("onnx is not installed, use \"pip install onnx==1.6.0\".")
         return
     print("Now translating model from onnx to paddle.")
 
     from x2paddle.op_mapper.onnx_op_mapper import ONNXOpMapper
     from x2paddle.decoder.onnx_decoder import ONNXDecoder
     from x2paddle.optimizer.onnx_optimizer import ONNXOptimizer
-
+    import onnxruntime
     model = ONNXDecoder(model_path)
     mapper = ONNXOpMapper(model, save_dir)
     optimizer = ONNXOptimizer(mapper)

x2paddle/decoder/onnx_decoder.py

@@ -316,12 +316,14 @@ class ONNXDecoder(object):
             model.ir_version, model.opset_import[0].version))
         if model.opset_import[0].version < 9:
             _logger.warning(
-                'Now, onnx2paddle main support convert onnx model opset_verison == 9,'
+                'Now, onnx2paddle support convert onnx model opset_verison == 9,'
                 'opset_verison of your onnx model is %d < 9,'
-                'some operator may cannot convert.',
+                'some operator maybe unsuccessful in convertion.',
                 model.opset_import[0].version)
 
         check_model(model)
+        self.check_model_running_state(onnx_model)
+
         model = onnx.shape_inference.infer_shapes(model)
         model = self.optimize_model_skip_op_for_inference(model)
         model = self.optimize_model_strip_initializer(model)
@@ -471,7 +473,44 @@ class ONNXDecoder(object):
             raise ValueError('name should not be empty')
         for s in ' .*?\\/-:':
             name = name.replace(s, '_')
-        return '_' + name
+        return 'x2paddle_' + name
+
+    def check_model_running_state(self, model_path):
+        try:
+            import onnxruntime as rt
+            version = rt.__version__
+            if version != '1.0.0':
+                print("onnxruntime==1.0.0 is required")
+                return
+        except:
+            raise Exception(
+                "onnxruntime is not installed, use \"pip install onnxruntime==1.0.0\"."
+            )
+
+        model = onnx.load(model_path)
+        model = onnx.shape_inference.infer_shapes(model)
+        if len(model.graph.value_info) < len(model.graph.node) - 1:
+            print(
+                "shape inference for some operators failed, those operators will be assignd node.out_shape==None, refer to https://github.com/onnx/onnx/blob/master/docs/ShapeInference.md"
+            )
+        try:
+            datatype_map = {
+                'tensor(int64)': 'int',
+                'tensor(float)': 'float32',
+                'tensor(int32)': 'int32'
+            }
+            input_dict = {}
+            sess = rt.InferenceSession(model_path)
+            for ipt in sess.get_inputs():
+                datatype = datatype_map[ipt.type]
+                input_dict[ipt.name] = np.random.random(
+                    ipt.shape).astype(datatype)
+
+            res = sess.run(None, input_feed=input_dict)
+        except:
+            raise Exception(
+                "onnxruntime inference onnx model failed, Please confirm the correctness of onnx model by onnxruntime, if onnx model is correct, please submit issue in github."
+            )
 
     def standardize_variable_name(self, graph):
         """

x2paddle/onnx_infer.py

-import os
-import sys
-import numpy as np
-import onnx
-import json
-import argparse
-from six import text_type as _text_type
-
-
-def arg_parser():
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--save_dir",
-                        "-s",
-                        type=_text_type,
-                        default=None,
-                        help="define save_dir")
-    return parser
-
-
-def main():
-    try:
-        import onnxruntime as rt
-        version = rt.__version__
-        if version != '0.4.0':
-            print("onnxruntime==0.4.0 is required")
-            return
-    except:
-        print(
-            "onnxruntime is not installed, use \"pip install onnxruntime==0.4.0\"."
-        )
-        return
-    parser = arg_parser()
-    args = parser.parse_args()
-
-    save_dir = args.save_dir
-    model_dir = os.path.join(save_dir, 'onnx_model_infer.onnx')
-
-    model = onnx.load(model_dir)
-    sess = rt.InferenceSession(model_dir)
-
-    inputs_dict = {}
-    for ipt in sess.get_inputs():
-        data_dir = os.path.join(save_dir, ipt.name + '.npy')
-        inputs_dict[ipt.name] = np.load(data_dir, allow_pickle=True)
-    res = sess.run(None, input_feed=inputs_dict)
-    for idx, value_info in enumerate(model.graph.output):
-        np.save(os.path.join(save_dir, value_info.name), res[idx])
-
-
-if __name__ == "__main__":
-    main()

x2paddle/op_mapper/onnx_op_mapper.py

@@ -32,7 +32,7 @@ import math
 import os
 import shutil
 from functools import reduce
-
+import onnxruntime as rt
 _logger = _logging.getLogger(__name__)
 
 
@@ -71,6 +71,7 @@ class ONNXOpMapper(OpMapper):
         self.used_custom_layers = dict()
         self.is_inference = False
         self.tmp_data_dir = os.path.join(save_dir, 'tmp_data')
+        self.tmp_outputs_dict = {}
         self.get_output_shapes()
 
         if not self.op_checker():
@@ -119,7 +120,7 @@ class ONNXOpMapper(OpMapper):
     def get_results_of_inference(self, model, value_infos, data_nodes):
         if not os.path.exists(self.tmp_data_dir):
             os.makedirs(self.tmp_data_dir)
-
+        inputs_dict = {}
         for data_node in data_nodes:
             value_info = value_infos[data_node]
             shape = value_info['shape']
@@ -129,34 +130,32 @@ class ONNXOpMapper(OpMapper):
                 if dim_shape == 0 and i != 0:
                     assert 'shape of input is not assigned'
             ipt = np.random.random(shape).astype(value_info['dtype'])
-            np.save(os.path.join(self.tmp_data_dir, data_node), ipt)
+            inputs_dict[data_node] = ipt
 
         model = onnx.shape_inference.infer_shapes(model)
         outputs = []
+
         for value_info in model.graph.value_info:
-            outputs.append(value_info)
+            outputs.append(value_info.name)
 
         model.graph.ClearField('output')
-        model.graph.output.MergeFrom(outputs)
+        model.graph.output.MergeFrom(model.graph.value_info)
         onnx.save(model, os.path.join(self.tmp_data_dir,
                                       'onnx_model_infer.onnx'))
+        sess = rt.InferenceSession(
+            os.path.join(self.tmp_data_dir, 'onnx_model_infer.onnx'))
+        res = sess.run(None, input_feed=inputs_dict)
+        self.tmp_outputs_dict = dict(zip(outputs, res))
 
-        is_success = os.system('onnx_infer --save_dir=' + self.tmp_data_dir)
-        if is_success != 0:
-            raise Exception("onnxruntime inference onnx model failed, Please \
-                             confirm the correctness of onnx model by onnxruntime, \
-                             if onnx model is valid, you can submit issue in github."
-                            )
         return
 
     def get_dynamic_shape(self, layer):
         """
         get dynamic shape from infer_result
         """
-        path = os.path.join(self.tmp_data_dir, layer + '.npy')
-        if not os.path.exists(path):
+        if layer not in self.tmp_outputs_dict:
             return [None, None, None]
-        output = np.load(path)
+        output = self.tmp_outputs_dict[layer]
         return output.tolist(), output.dtype, output.shape
 
     def get_output_shapes(self):